July 2026 · 1,599 commits since February
RELEASES
Tracking the Evolution of OpenMake LLM
Grouped by month, straight from the commit history. The package version sat at 1.5.6 from February through June; release-please landed in July and it now advances with every release (currently 1.7.0).
July 2026294 commitsLATEST
Git-native agent tasks, NotebookLM grounding, Cowork’s local execution bridge to the desktop app, and automated release versioning
architecture
Agent Autonomy
- Git Task WorkflowAn agent task can connect GitHub, clone a repository, edit it, capture a diff, and carry it through commit, push, and pull request — with per-file collapsible diffs, inline rendering in chat, copy, and .patch download
- Sub-Agents & Cross-Task LearningPhase 5 added delegated sub-agents, learning carried across tasks, extended verification, and embedding-driven dynamic tool selection
- Templates, Schedules & CheckpointsPhase 6 turned one-off runs into a product surface: reusable task templates, recurring schedules with run history and failure alerts, and mid-turn checkpoints
- Approval Modes & SteeringThree approval modes — Manual, Auto, and Skip — plus mid-run steering so you can redirect a task without restarting it
- Cowork — Local Execution BridgeAgent tasks can now run tools on your own machine instead of the server sandbox: a TaskExecutor abstraction, a local bridge runner shared with the desktop app, and a composer toggle with a badge showing which tasks are running locally
architecture
Knowledge & Memory
- NotebookLM GroundingPin one of your Google NotebookLM notebooks as conversation context, with a picker in the composer and catalog-managed tool allowlisting
- Procedural SkillsA successful run can be captured as a replayable procedural skill rather than being repeated by hand
- Semantic Memory in Agent TasksThe three-tier memory system was wired into agent tasks so a run can draw on what earlier conversations established
misc
Integrations & Surface
- Artifacts Over the APIThe OpenAI-compatible API and the Discord bot both return artifacts, with optional URL publishing
- Desktop ShellThe Electron app grew from a bare shell into the Cowork local bridge runner, with its own release track (v1.0.0 → v1.2.1) and a server-manifest-based self-updater for its unsigned macOS build
- Admin Schedule OversightAdministrators can review the recurring agent-task schedules of every user in one table
- Domain MoveThe public deployment moved to chat.openmake.cc, with guest UX, resizable artifacts, and a rewritten README
security
Release Hygiene
- Automated Versioningrelease-please was adopted for Conventional-Commits-based semantic versioning, so the package version now advances every release (1.5.7 → 1.6.0 → 1.7.0) instead of standing still
- Credential CleanupPersonal identifiers and hardcoded credentials were removed from the now-public repository
435 Files Modified
June 2026402 commits
The autonomous agent arrives — a persistent Docker sandbox with browser automation, self-directed planning, and live progress
architecture
Agent Tasks
- Persistent Sandbox & Approval GatesPhase 1 of the Manus-style agent: a long-lived Docker sandbox with shell, Python, and file tools, fronted by human-in-the-loop approval
- Browser AutomationPlaywright Chromium inside the container, with a hard cap on browser calls so a run cannot spiral into unbounded crawling
- Autonomous Planning & ResumeThe agent decomposes a goal, re-plans when it stalls, and can resume from a checkpoint instead of starting over
- Live Computer PanelPlan, steps, and outputs stream to a live panel over WebSocket while the task runs
misc
Outputs & Notifications
- Excel & PDF DeliverablesSandbox runs can produce real documents, not just text
- Artifacts From Task OutputResults are extracted and rendered as artifacts so the work is visible rather than buried in a log
- Completion AlertsIn-app notifications plus web-push, so a finished run reaches you even with the browser closed
security
Tool Surface
- Filesystem & Review ToolsFilesystem MCP tools landed alongside security review, code review, and planning tools built on shared analyzer patterns
- Per-Sandbox Tool AllowlistSandbox tasks declare an explicit built-in tool allowlist, with the approval gate applied on top
misc
Admin
- MCP Catalog in Admin TabsThe MCP catalog and monitoring pages were folded into the admin tab surface
- Agent Prompt SuggestionsModel-proposed agent prompts now go through an admin approval workflow
1166 Files Modified
May 2026460 commits
Collapsing to a single local model, adding a role registry to route around it, and consolidating the admin surface
architecture
Model Simplification
- Brand Profiles RemovedPer-brand model profiles were deleted, and the model selector, profile resolver, strategy executor, and model resolver were all simplified for a single local model
- Model Selector Moved to SettingsModel choice left the chat header and became a settings concern
- Startup Model ValidationThe server checks at boot that the configured models actually exist, so a typo fails immediately rather than on first message
architecture
Role Orchestration
- Central Role RegistryA single registry resolves a model per functional role with an explicit fallback chain, laying the groundwork for later per-user and org-wide mappings
misc
Agent & Skill Ingest
- Import From GitAgents and skills can be imported straight from a Git repository, arriving as drafts with an approval status
security
Admin & Compliance
- Admin ConsolidationFive separate admin pages were unified under section tabs
- Audit DashboardPagination plus an integrated alert-history view
- GDPR WorkConsent hold for users under 14, and a modal spelling out the impact before an account is deleted
512 Files Modified
April 202661 commits
A deliberate security and infrastructure month — CSRF, CSP, SSRF guards, and a pluggable key-value store
security
Security Hardening
- CSRF Double-SubmitDouble-submit cookie middleware shipped in warn mode first, with the frontend API client injecting the token on every mutation
- CSP Tighteningscript-src-attr enforced through a SHA-256 hash whitelist instead of blanket inline permission
- Permissions-PolicyCamera, microphone, and geolocation denied outright
- Centralised GuardsSSRF, WebSocket, blacklist, and cookie rules moved into one security config, with an explicit opt-out for insecure cookies in non-TLS deployments
data
Storage & Scale
- Pluggable Key-Value StoreA KeyValueStore interface with memory and Redis implementations, selected by STORAGE_BACKEND
- Multi-Instance ReadinessOAuth state moved off process memory and onto the shared store
stability
Runtime
- Container PackagingDockerfile and docker-compose added for reproducible deployment
- Health & Circuit BreakerModel health monitoring with a circuit breaker and an admin view
- Versioned Prompt TemplatesA database schema and repository for prompt template versioning
127 Files Modified
March 2026214 commits
Harness engineering — verification, context compaction, and pulling every hard-coded value out of the code
architecture
Execution Quality
- Confidence-Gated StrategiesStrategy selection gated on confidence, with post-execution verification and an informed fallback when a run does not hold up
- Generate-VerifyThe legacy agent-to-agent path was removed in favour of a Generate-Verify loop
performance
Context Engineering
- Semantic CompactionLong histories are compacted by meaning rather than truncated by position
- Dynamic Token BudgetsBudgets computed per request instead of a fixed ceiling, with tool-result compaction and just-in-time memory filtering
- Context Garbage CollectionA dedicated service reclaims context that no longer earns its place
testing
Observability
- Evaluation Pipeline & Trace AnalyzerAn evaluation pipeline and trace analyzer landed with supporting test mocks, and Generate-Verify quality metrics are persisted
misc
No-Hardcoding
- Configuration ExternalisedPrompts, keywords, regex patterns, thresholds, and timers moved out of the code into config files across five phases
- Record Maps Over SwitchSwitch-case dispatch replaced with record maps, and magic numbers promoted to named constants
548 Files Modified
February 2026168 commits
The first public shape of the platform — retrieval, deep research, and the MCP tool surface
architecture
Core Features
- Retrieval & Deep ResearchA retrieval system with its backend services, routes, tests, and UI, followed by deep research with external service and pricing configuration
- Documents & OfflineDocument management plus a service worker for offline behaviour
architecture
Tooling
- MCP Tool TiersTool tiering and routing logic, with sequential_thinking exempted from tier limits — the tier model itself was retired later in the year
misc
Language & Prompts
- Language PolicyA language policy layer and a restructured agent prompt system, with dynamic response language enabled by default
security
Security
- Upload HardeningOptional authentication on upload routes and stricter file upload validation
1303 Files Modified